LOCALLY MONOTONE BOOLEAN AND PSEUDO-BOOLEAN FUNCTIONS
MIGUEL COUCEIRO, JEAN-LUC MARICHAL, AND TAM ´AS WALDHAUSER
Abstract. We propose local versions of monotonicity for Boolean and pseudo- Boolean functions: say that a pseudo-Boolean (Boolean) function isp-locally monotone if none of its partial derivatives changes in sign on tuples which differ in less thanppositions. As it turns out, this parameterized notion provides a hierarchy of monotonicities for pseudo-Boolean (Boolean) functions.
Local monotonicities are shown to be tightly related to lattice counter- parts of classical partial derivatives via the notion of permutable derivatives.
More precisely,p-locally monotone functions are shown to havep-permutable lattice derivatives and, in the case of symmetric functions, these two notions coincide. We provide further results relating these two notions, and present a classification ofp-locally monotone functions, as well as of functions havingp- permutable derivatives, in terms of certain forbidden “sections”, i.e., functions which can be obtained by substituting constants for variables. This description is made explicit in the special case whenp=2.
1. Introduction
Throughout this paper, let[n]={1, . . . , n}andB={0,1}. We are interested in the so-called Boolean functionsf∶Bn→Band pseudo-Boolean functionsf∶Bn→R, wheren denotes the arity off. The pointwise ordering of functions is denoted by
≤, i.e., f ≤ g means that f(x) ≤ g(x) for all x ∈ Bn. The negation of x ∈ B is defined by x=x⊕1, where ⊕ stands for addition modulo 2. For x, y∈ B, we set x∧y=min(x, y)andx∨y=max(x, y).
Fork∈[n],x∈Bn, anda∈B, letxak be the tuple inBn whosei-th component is a, ifi=k, andxi, otherwise. We use the shorthand notationxabjkfor(xaj)bk=(xbk)aj. More generally, forS ⊆[n], a∈Bn, and x∈ BS, let axS be the tuple in Bn whose i-th component isxi, ifi∈S, andai, otherwise.
Leti∈[n]and f∶Bn →R. A variable xi is said to be essential in f, or that f depends onxi, if there existsa∈Bn such thatf(a0i)≠f(a1i). Otherwise,xi is said to beinessential in f. Let S ⊆[n]and f∶Bn →R. We say that g∶BS → Ris an S-section of f if there exists a∈ Bn such that g(x)=f(axS)for all x∈BS. By a section of f we mean anS-section of f for someS ⊆[n], i.e., any function which can be obtained fromf by replacing some of its variables by constants.
The(discrete) partial derivative off∶Bn→Rwith respect to itsk-th variable is the function ∆kf∶Bn→Rdefined by ∆kf(x)=f(x1k)−f(x0k); see [8, 12]. Note that
Date: March 27, 2012.
2010Mathematics Subject Classification. 06E30, 94C10.
Key words and phrases. Boolean function, pseudo-Boolean function, local monotonicity, dis- crete partial derivative, join and meet derivatives.
1
∆kf does not depend on itsk-th variable, hence it could be regarded as a function of arityn−1, but for notational convenience we define it as ann-ary function.
A pseudo-Boolean functionf∶Bn→Rcan always be represented by a multilinear polynomial of degree at mostn(see [13]), that is,
(1) f(x) = ∑
S⊆[n]
aS ∏
i∈S
xi,
whereaS∈R. For instance, the multilinear expression for a binary pseudo-Boolean function is given by
(2) a0+a1x1+a2x2+a12x1x2.
This representation is very convenient for computing the partial derivatives of f. Indeed, ∆kf can be obtained by applying the corresponding formal derivative to the multilinear representation off. Thus, from (1), we immediately obtain
(3) ∆kf(x) = ∑
S∋k
aS ∏
i∈S∖{k}
xi.
We say that f is isotone (resp. antitone) in its k-th variable if ∆kf(x) ≥ 0 (resp. ∆kf(x) ≤ 0) for all x ∈ Bn. If f is either isotone or antitone in its k-th variable, then we say thatf ismonotone in its k-th variable. Iff is isotone (resp.
antitone, monotone) in all of its variables, then f is an isotone (resp. antitone, monotone) function.1 It is clear that any section of an isotone (resp. antitone, monotone) function is also isotone (resp. antitone, monotone). Thus defined, a functionf∶Bn→Ris monotone if and only if none of its partial derivatives changes in sign onBn.
Noteworthy examples of monotone functions include the so-calledpseudo-polyno- mial functions [2, 3] which play an important role, for instance, in the qualita- tive approach to decision making; for general background see, e.g., [1, 6]. In the current setting, pseudo-polynomial functions can be thought of as compositions p○(φ1, . . . , φn) of (lattice) polynomial functions p∶[a, b]n → [a, b], a < b, with unary functions φi∶B→ [a, b], i∈ [n]. Interestingly, pseudo-polynomial functions f∶Bn→Rcoincide exactly with those pseudo-Boolean functions that are monotone.
Theorem 1. A pseudo-Boolean function is monotone if and only if it is a pseudo- polynomial function.
Proof. Clearly, every pseudo-polynomial function is monotone. For the converse, suppose that f∶Bn → R is monotone and let a ∈ R be the minimum and b ∈ R the maximum off. Constant functions are obviously pseudo-polynomial functions, therefore we assume a < b. Define φi∶B → {a, b} by φi(0) = a and φi(1) = b if f is isotone in its i-th variable and φi(0) = b and φi(1) = a otherwise. Let p∶{a, b}n → [a, b] be given by p = f ○(φ−11, . . . , φ−n1). Thus defined, p is isotone (i.e., order-preserving) in each variable and hence, by Theorem D in Goodstein [10, p. 237], there exists a polynomial functionp′∶[a, b]n→[a, b]such thatp′∣{a,b}n=p.
Thereforef is the pseudo-polynomial functionp′○(φ1, . . . , φn). In the special case of Boolean functions, monotone functions are most frequent among functions of small (essential) arity. For instance, among binary functions
1Note that the terms “positive” and “nondecreasing” (resp. “negative” and “nonincreasing”) are often used instead of isotone (resp. antitone), and it is also customary to use the word “mono- tone” only for isotone functions.
f∶B2→B, there are exactly two non-monotone functions, namely the Boolean sum x1⊕x2 and its negation x1⊕x2⊕1. Each of these functions is in fact highly non-monotone in the sense that any of its partial derivatives changes in sign when negating its unique essential variable; this is not the case, e.g., withf(x1, x2, x3)= x1−x1x2+x2x3 which is non-monotone but none of its partial derivatives changes in sign when negating any of its variables (see Example 6 below).
This fact motivates the study of these “skew” functions, i.e., these highly non- monotone functions. To formalize this problem we propose the following param- eterized relaxations of monotonicity: a function f∶Bn → R is p-locally monotone if none of its partial derivatives changes in sign when negating less than p of its variables, or equivalently, on tuples which differ in less than p positions. With this terminology, our problem reduces to asking which Boolean functions are not 2- locally monotone. As we will see (Corollary 10), these are precisely those functions that have the Boolean sum or its negation as a binary section.
In this paper we extend this study to pseudo-Boolean functions and show that these parameterized relaxations of monotonicity are tightly related to the following lattice versions of partial derivatives. For f∶Bn →Rand k∈[n], let∧kf∶Bn →R and∨kf∶Bn→Rbe thepartial lattice derivatives defined by
∧kf(x) = f(x0k)∧f(x1k) and ∨kf(x) = f(xk0)∨f(x1k).
The latter, known as thek-thjoin derivative off, was proposed by Fadini [9] while the former, known as the k-th meet derivative of f, was introduced by Thayse [16]. In [17] these lattice derivatives were shown to be related to so-called prime implicants and implicates of Boolean functions which play an important role in the consensus method for Boolean and pseudo-Boolean functions. For further back- ground and applications see, e.g., [4, 5, 7, 15, 18].
Observe that, just like in the case of the partial derivative ∆kf, thek-th variable of each of the lattice derivatives∧kf and∨kf is inessential.
The following proposition assembles some basic properties of lattice derivatives.
Proposition 2. For any pseudo-Boolean functions f, g∶Bn → R and j, k ∈ [n], j≠k, the following hold:
(i) ∧k∧kf=∧kf and∨k∨kf =∨kf; (ii) if f≤g, then∧kf ≤∧kg and∨kf ≤∨kg;
(iii) ∧j∧kf =∧k∧jf and∨j∨kf =∨k∨jf; (iv) ∨k∧jf ≤∧j∨kf.
From equations (1) and (3) it follows that every function is (up to an additive constant) uniquely determined by its partial derivatives. As it turns out, this does not hold when lattice derivatives are considered. However, as we shall see (Theorem 22), there are only two types of such exceptions.
Now, if ann-ary pseudo-Boolean function is 2-locally monotone, then for every j, k∈[n], j≠k, we have∨k∧jf =∧j∨kf (see Lemma 11 below). This motivates the notion of permutable lattice derivatives. As it turns out,p-local monotonicity off implies permutability ofpof its lattice derivatives (see Theorem 21). However the converse does not hold (see Example 24).
The structure of this paper goes as follows. In Section 2 we formalize the notion of p-local monotonicity and show that it gives rise to a hierarchy of monotonici- ties whose largest member is the class of all n-ary pseudo-Boolean functions (this
is the case when p= 1) and whose smallest member is the class of n-ary mono- tone functions (this is the case whenp=n). We also provide a characterization of p-locally monotone functions in terms of “forbidden” sections; as mentioned, this characterization is made explicit in the special case when p= 2. In Section 3 we introduce the notion of permutable lattice derivatives. Similarly to local mono- tonicity, the notion of permutable lattice derivatives gives rise to nested classes, each of which is also described in terms of its sections. In the Boolean case and for p=2, these two parameterized notions are shown to coincide; this does not hold for pseudo-Boolean functions even whenp=2 (see Example 13). (At the end of Sec- tion 3 we also provide some game-theoretic interpretations ofp-local monotonicity andp-permutability of lattice derivatives.) However, in Section 4, we show that a symmetric function isp-locally monotone if and only if it hasp-permutable lattice derivatives. In the last section we discuss directions for future research.
2. Local monotonicities
The following definition formulates a local version of monotonicity given in terms of Hamming distance between tuples. In what follows we assume thatp∈[n]. Definition 3. We say that f∶Bn →R is p-locally monotone if, for everyk ∈[n] and everyx,y∈Bn, we have
∑
i∈[n]∖{k}
∣xi−yi∣ < p ⇒ ∆kf(x)∆kf(y) ≥ 0.
Anyp-locally monotone pseudo-Boolean function is alsop′-locally monotone for every p′ ≤p. Every function f∶Bn → R is 1-locally monotone, and f is n-locally monotone if and only if it is monotone. Thusp-local monotonicity is a relaxation of monotonicity, and the nested classes ofp-locally monotone functions forp=1, . . . , n provide a hierarchy of monotonicities for n-ary pseudo-Boolean functions. The weakest nontrivial condition is 2-local monotonicity, therefore we will simply say that f is locally monotone whenever f is 2-locally monotone.2 If f is p-locally monotone for somep<n but not(p+1)-locally monotone, then we say thatf is exactly p-locally monotone, or that thedegree of local monotonicity of f isp.
If f∶Bn → B is a Boolean function, then ∆kf(x) ∈ {−1,0,1} for all x ∈ Bn, hence the condition ∆kf(x)∆kf(y)≥0 in the definition ofp-local monotonicity is equivalent to
(4) ∣∆kf(x)−∆kf(y)∣ ≤ 1.
From this it follows that a Boolean functionf∶Bn→B is locally monotone if and only if
(5) ∣∆kf(x)−∆kf(y)∣ ≤ ∑
i∈[n]∖{k}∣xi−yi∣.
(see [14, Lemma 5.1] for a proof of (5) in a slightly more general framework). In a sense, the latter identity means that ∆kf is “1-Lipschitz continuous”.
The following proposition is just a reformulation of the definition ofp-local mono- tonicity.
2In [11], local monotonicity is used to refer to Boolean functions which are monotone (i.e., isotone or antitone in each variable).
Proposition 4. A function f∶Bn → R is p-locally monotone if and only if, for every k∈[n],S⊆[n]∖{k}, with ∣S∣=p−1, and everya∈Bn,x,y∈BS, we have (6) ∆kf(axS)∆kf(ayS) ≥ 0.
Equivalently, a pseudo-Boolean function isp-locally monotone if and only if none of its partial derivatives changes in sign when negating less thanpof its variables.
As a special case, we have thatf∶Bn→Ris locally monotone if and only if, for everyj, k∈[n],j≠k, and everyx∈Bn, we have
(7) ∆kf(x0j)∆kf(x1j) ≥ 0.
Equivalently, a pseudo-Boolean function is locally monotone if and only if none of its partial derivatives changes in sign when negating any of its variables.
By (4) we see that, for Boolean functions f∶Bn → B, inequality (7) can be replaced with∣∆jkf(x)∣≤1, where ∆jkf(x)=∆j∆kf(x)=∆k∆jf(x).
Example 5. As observed, the binary Boolean sum
f1(x1, x2) = x1⊕x2 = x1+x2−2x1x2
and the binary Boolean equivalence
f2(x1, x2) = f1(x1, x2) = x1⊕x2⊕1 = 1−x1−x2+2x1x2
are not locally monotone. Indeed, we have∣∆12f1(x1, x2)∣=∣∆12f2(x1, x2)∣=2.
Example 6. Consider the ternary Boolean functionf∶B3→Bgiven by f(x1, x2, x3) = x1−x1x2+x2x3.
Since ∆2f may change in sign (∆2f(x)=x3−x1), the functionf is not monotone.
However,f is locally monotone since∣∆12f(x)∣=1,∣∆13f(x)∣=0, and∣∆23f(x)∣= 1. Thus f is exactly 2-locally monotone. Example 26 in Section 4 provides, for eachp≥2, examples of exactlyp-locally monotone functions.
Fact 7. A functionf∶Bn→Risp-locally monotone if and only if so isαf+β for every α, β ∈ R, with α≠0. The same holds for any function obtained from f by negating some of its variables.
The next theorem gives a characterization of p-locally monotone functions in terms of their sections.
Theorem 8. A functionf∶Bn→Risp-locally monotone if and only if everyp-ary section of f is monotone.
Proof. We just need to observe that the inequality (6) is equivalent to ∆kg(x)∆kg(y)≥ 0, wheregis thep-ary section off defined byg(x)=f(axS∪{k}), whereSis a(p−1)- subset of[n]∖{k}. Thusf isp-locally monotone if and only if ∆kg(x)∆kg(y)≥0 holds for everyx,y∈BS∪{k}, and for everyS∪{k}-sectiong off. By combining (7) with Theorem 8, we can easily verify the following corollary.
Corollary 9. A functionf∶Bn→Ris locally monotone if and only if every binary section (2) of f satisfiesa1(a1+a12)≥0and a2(a2+a12)≥0.
Since every binary Boolean function is monotone except forx⊕y andx⊕y⊕1, we also obtain the following corollary.
Corollary 10. A Boolean function f∶Bn → B is locally monotone if and only if neitherx⊕y nor x⊕y⊕1 is a section off.
3. Permutable lattice derivatives
The aim of this section is to relate commutation of lattice derivatives to p- local monotonicity. The starting point is the characterization of locally monotone Boolean functions given in Theorem 12 below.
Lemma 11. Iff∶Bn→Ris locally monotone, then∨k∧jf =∧j∨kf for allj, k∈[n], j≠k.
Proof. Let f∶Bn → R be a locally monotone function, and let j, k ∈ [n], j ≠ k.
Setting a= f(x00jk), b= f(x01jk), c =f(x10jk), andd =f(x11jk), the desired equality
∨k∧jf(x)=∧j∨kf(x)takes the form
(8) (a∧c)∨(b∧d) = (a∨b)∧(c∨d).
Since f is 2-locally monotone, the binary section g(u, v) = f(xuvjk) is monotone, according to Theorem 8. If g is isotone in u, then a ≤c and b ≤ d, while if g is antitone inu, thena≥candb≥d. Similarly, we have eithera≤bandc≤dora≥b and c ≥d, depending on whether g is isotone or antitone in v. Thus we need to consider four cases, and in each one of them it is straightforward to verify (8).
Theorem 12. A Boolean function f∶Bn → B is locally monotone if and only if
∨k∧jf=∧j∨kf holds for allj, k∈[n],j≠k.
Proof. Iff is locally monotone, then∨k∧jf =∧j∨kf by Lemma 11. If f is not locally monotone, then Corollary 10 implies that there existsa∈Bn andj, k∈[n], j≠k, such that the binary sectiong(u, v)=f(ajkuv)is of the formg(u, v)=u⊕v or g(u, v)=u⊕v⊕1. Then we have
∨k∧jf(a) = (g(0,0)∧g(1,0))∨(g(0,1)∧g(1,1))=0,
∧j∨kf(a) = (g(0,0)∨g(0,1))∧(g(1,0)∨g(1,1))=1,
showing that∨k∧jf≠∧j∨kf.
As the next example shows, Theorem 12 is not valid for pseudo-Boolean func- tions.
Example 13. Letf be the binary pseudo-Boolean function defined byf(0,0)=1, f(0,1)=4, f(1,0)=2 and f(1,1)=3. Then we have ∨2∧1f =∧1∨2f =3 and
∨1∧2f=∧2∨1f=2. However,f is not locally monotone since ∆1f(x02)∆1f(x12)=
−1.
Lemma 11 and Theorem 12 motivate the following notion of permutability of lattice derivatives, and its relation to local monotonicities.
Definition 14. We say that a pseudo-Boolean functionf∶Bn→Rhasp-permutable lattice derivatives if, for every p-subset{k1, . . . , kp}⊆[n], every choice of the op- eratorsOki ∈{∧ki,∨ki} (i=1, . . . , p), and every permutationπ∈Sp, the following identity holds:
Ok1⋯Okpf = Okπ(1)⋯Okπ(p)f.
If f∶Bn → R hasn-permutable lattice derivatives, then we simply say that f has permutable lattice derivatives.
Every function f∶Bn → R has 1-permutable lattice derivatives. We will see in Theorem 23 that if a functionf∶Bn→Rhasp-permutable lattice derivatives, then it also hasp′-permutable lattice derivatives for every p′≤p.
Fact 15. A functionf∶Bn→Rhasp-permutable lattice derivatives if and only if so hasαf+β for everyα, β∈R, withα≠0. The same holds for any function obtained fromf by negating some of its variables.
Fact 16. A function f∶Bn→R has p-permutable lattice derivatives if and only if every p-ary section of f has permutable lattice derivatives.
In the particular case whenp=2, we have the following description of functions having 2-permutable lattice derivatives. The proof is a straightforward verification of cases.
Proposition 17. A functionf∶Bn→Rhas2-permutable lattice derivatives if and only if every binary section (2) of f satisfies a1a12 ≥ 0 or a2a12 ≥ 0 or ∣a12∣ ≤
∣a1∣∨∣a2∣.
Lemma 11 shows that the class of 2-locally monotone pseudo-Boolean functions is a subclass of that of pseudo-Boolean functions which have 2-permutable lattice derivatives. Example 13 then shows that this inclusion is strict. On the other hand, according to Theorem 12, a Boolean function is 2-locally monotone if and only if it has 2-permutable lattice derivatives. Example 24 below shows that the analogous equivalence does not hold for p > 2. However, p-local monotonicity implies p-permutability of lattice derivatives of any pseudo-Boolean function (see Theorem 21 below). To this extent, let us first study how the degree of local monotonicity is affected by taking lattice derivatives.
Lemma 18. If f∶Bn →Ris monotone, then ∧jf and ∨jf are also monotone for allj∈[n].
Proof. Clearly, iff is monotone, then so arefj0(x)=f(x0j)andfj1(x)=f(x1j), for allj∈[n]. Moreover, iff is isotone (resp. antitone) inxk, then bothfj0andfj1are also isotone (resp. antitone) in xk. Since∧ and ∨ are isotone functions, we have that for everyj∈[n], both∧jf(x)=fj0(x)∧fj1(x)and∨jf(x)=fj0(x)∨fj1(x)are
monotone.
Theorem 19. If f∶Bn →R is p-locally monotone, then ∧jf and ∨jf are (p−1)- locally monotone for all j∈[n].
Proof. Suppose thatf∶Bn→Risp-locally monotone. By Theorem 8, it suffices to show that all(p−1)-ary sections of∧jf and ∨jf are monotone. We consider only
∨jf, the other case can be dealt with in a similar way.
Lethbe a (p−1)-ary section of∨jf defined byh(x)=∨jf(axS)for all x∈BS, where a ∈Bn and S ⊆[n]is a (p−1)-subset. Let T = S∪{j}, and let us define g∶BT →Rbyg(y)=f(ayT)for ally∈BT. Clearly,gis a section off, and the arity of g is either p−1 or p, depending on whether j belongs toS or not. A simple calculation shows that h(y∣S) = ∨jg(y) for all y ∈ BT, where y∣S stands for the restriction of y to S. This means that if j∉ S, thenh can be obtained from∨jg by deleting its inessentialj-th variable, and h=∨jg ifj ∈S. Since f is p-locally monotone, g is monotone by Theorem 8, thus we can conclude with the help of
Lemma 18 thathis monotone as well.
Corollary 20. Let 0 ≤ ℓ <p ≤n. If f∶Bn → R is p-locally monotone, then, for every ℓ-subset {k1, . . . , kℓ}⊆[n] and every choice of the operators Oki∈{∨ki,∧ki} (i=1, . . . , ℓ), the functionOk1⋯Okℓf is(p−ℓ)-locally monotone. In particular, if ℓ≤p−2, thenOk1⋯Okℓf is locally monotone.
Remark 1. We will see in Example 26 of Section 4 that Theorem 19 cannot be sharpened, i.e., the lattice derivatives of a p-locally monotone function are not necessarilyp-locally monotone, not even in the case of Boolean functions.
With the help of Corollary 20 we can now prove the promised implication between p-local monotonicity and p-permutability of lattice derivatives, thus generalizing Lemma 11.
Theorem 21. Iff∶Bn→Risp-locally monotone, then it hasp-permutable lattice derivatives.
Proof. Let f∶Bn → R be a p-locally monotone function, let {k1, . . . , kp} be a p- subset of[n], and letOki∈{∧ki,∨ki}fori=1, . . . , p. We need to show that for any permutationπ∈Sp the following identity holds:
Ok1⋯Okpf = Okπ(1)⋯Okπ(p)f.
SinceSpis generated by transpositions of the form(i i+1), it suffices to prove that Ok1⋯Oki−1OkiOki+1Oki+2⋯Okpf = Ok1⋯Oki−1Oki+1OkiOki+2⋯Okpf, and for this it is sufficient to verify that
(9) OkiOki+1g = Oki+1Okig,
where g stands for the functionOki+2⋯Okpf. From Corollary 20 it follows thatg is locally monotone, and then Lemma 11 proves (9) if one of Oki, Oki+1 is a meet and the other is a join derivative. (If both are meet or both are join, then (9) is
trivial.)
A natural question regarding lattice derivatives is whether a function can be re- constructed from its derivatives. As the next theorem shows, the answer is positive for almost all functions.
Theorem 22. Letf, g∶Bn→Rbe pseudo-Boolean functions such that for allk∈[n] we have ∨kf =∨kg and∧kf =∧kg. Then eitherf =g or there exists a one-to-one function α∶B→Rsuch that f(x)=α(x1⊕ ⋯ ⊕xn) andg(x)=α(x1⊕ ⋯ ⊕xn⊕1) for allx∈Bn.
Proof. To make the proof more vivid, we present it through the analysis of the following game. Alice picks a secret function f∶Bn →R, and Bob tries to identify this function by asking the values of its lattice derivatives. If he can do this, then he wins, otherwise Alice is the winner. We show that Bob has a winning strategy unlessf is a function of the special form in the statement of the theorem.
Let us regardBn as the set of vertices of then-dimensional cube, and let Alice write the values of f to the corresponding vertices. Now the possible winning strategy for Bob is based on the following four basic observations.
1. Bob can determine the unordered pair of numbers written to the endpoints of any edge of the cube. Indeed, the endpoints of an edge are of the form x0k,x1k, and it is clear that{∧kf(x),∨kf(x)}={f(xk0), f(x1k)}.
2. If Bob can find the value of f at one point, then he can win. According to the previous observation, knowing the value at one vertex of the cube, Bob can figure out the values written to the neighboring vertices. Since the graph of the cube is connected, he can determine all values off this way.
3. If f(x0k)=f(x1k)for some x∈Bn,k∈[n], then Bob can win. This follows immediately from the first two observations.
4. If the range of f contains at least three elements, then Bob can win. We can suppose that the previous observation does not apply, i.e., for every edge Bob detects a two-element set. If f takes on at least three different values, then, by the connectedness of the cube, there exists a vertexxand two edges incident with this vertex such that the two-element setsE1 and E2 corresponding to these edges are different. Then E1∩E2 must be a one-element set3 containing the value of f(x), and then Bob can win as explained in the second observation.
From these observations we can conclude that Bob has a winning strategy unless the range off contains exactly two numbers andf(x0k)≠f(x1k), for allx∈Bn, k∈ [n]. This means thatf is of the following form for someu≠v∈R:
f(x) = {u , if ∣x∣ is even ; v , if ∣x∣ is odd ,
where∣x∣=∑ni=1xi. In other words,f(x)=α(x1⊕⋯⊕xn), whereα(0)=u, α(1)=v.
In this case Bob can determinef only up to interchanging uandv, i.e., he cannot distinguishf from g(x)=α(x1⊕ ⋯ ⊕xn⊕1), so he has only 50% chance to win.
(Indeed,f and g have the same lattice derivatives, namely their meet derivatives are all constantu∧v, while their join derivatives are all constantu∨v.) The following theorem shows that, as in the case of local monotonicity, the classes of functions having permutable lattice derivatives form a chain under inclusion.
Theorem 23. If f∶Bn →Rhas (p+1)-permutable lattice derivatives, then f has p-permutable lattice derivatives.
Proof. Letf∶Bn→Rbe a function that has(p+1)-permutable lattice derivatives.
Using the same notation as in the proof of Theorem 21, it suffices to prove that Ok1⋯Oki−1OkiOki+1Oki+2⋯Okpf = Ok1⋯Oki−1Oki+1OkiOki+2⋯Okpf.
Let g1 and g2 be the (n−p)-ary functions obtained from the left-hand side and from the right-hand side of this equality by deleting their inessential variables xk1, . . . , xkp. If Oki = ∧ki, Oki+1 = ∧ki+1 or Oki = ∨ki, Oki+1 = ∨ki+1, then g1 = g2
holds trivially. Let us now assume thatOki=∨ki, Oki+1=∧ki+1; the remaining case Oki=∧ki, Oki+1=∨ki+1 is similar.
By Proposition 2, we haveg1≤g2. Since the two (types of) functions given in Theorem 22 are order-incomparable, if g1 ≠ g2, then the lattice derivatives of g1
andg2cannot all coincide. Thus there existsj∈[n]∖{k1, . . . , kp}andOj∈{∧j,∨j} such that Ojg1 ≠Ojg2. Taking into account the definition of g1 and g2, we can rewrite this inequality as
OjOk1⋯Oki−1OkiOki+1Oki+2⋯Okpf ≠ OjOk1⋯Oki−1Oki+1OkiOki+2⋯Okpf, which contradicts the fact thatf has(p+1)-permutable lattice derivatives.
If f∶Bn → B is a Boolean function with p-permutable lattice derivatives for some p≥2, then f has 2-permutable lattice derivatives by Theorem 23, and then Theorem 12 implies that f is 2-locally monotone. Unfortunately, nothing more
3IfE1∩E2is empty, then Alice is cheating!
can be said about the degree of local monotonicity of a Boolean function withp- permutable lattice derivatives. Indeed, the next example shows that there exist n-ary Boolean functions with n-permutable lattice derivatives that are exactly 2- locally monotone.
Example 24. Letfn∶Bn→Bbe the function that takes the value 1 on all tuples of the form
x = (
³¹¹¹¹¹¹¹¹¹¹·¹¹¹¹¹¹¹¹¹µm
1, . . . ,1,0, . . . ,0) with 0≤m≤n,
and takes the value 0 everywhere else. Using Corollary 10, it is not difficult to verify that fn is 2-locally monotone. However, if n ≥ 3, then fn is not 3-locally monotone, since
∆2f(0,0,0,0, . . . ,0) = −1,
∆2f(1,0,1,0, . . . ,0) = 1.
Thusfn is exactly 2-locally monotone.
We will show by induction on n that fn has n-permutable lattice derivatives.
First we compute the meet derivatives
∧kfn(x)={1, ifx1=⋯=xk−1=1 andxk+1=⋯=xn=0 ; 0, otherwise.
Since∧kf takes the value 1 only at one tuple, it is monotone. The join derivative
∨kfnis essentially the same as the functionfn−1(up to the inessentialk-th variable of∨kfn), that is,
(10) ∨kfn(x) = fn−1(x1, . . . , xk−1, xk+1, . . . , xn).
Now it follows that if{k1, . . . , kn}=[n]andOki∈{∧ki,∨ki} (i=1, . . . , n), then (11) Ok1⋯Okn−1Oknf = Okπ(1)⋯Okπ(n−1)Oknf
holds for every permutation π ∈ Sn−1. (If Okn = ∧kn, then we use Theorem 21 and the fact that ∧knf is monotone, and if Okn =∨kn, then we use (10) and the induction hypothesis.) On the other hand, from the 2-local monotonicity of f we can conclude that
(12) Ok1⋯Okn−2Okn−1Oknf = Ok1⋯Okn−2OknOkn−1f
with the help of Theorem 12. SinceSn is generated bySn−1 and the transposition (n−1n), we see from (11) and (12) thatf hasn-permutable lattice derivatives.
We finish this section with game-theoretic interpretations of the parameterized notions of local monotonicity and permutability of lattice derivatives. Identifying Bn with the power set of[n], we can regard a pseudo-Boolean functionf∶Bn→R as a cooperative game, where [n] is the set of players and f(C) is the worth of coalitionC⊆[n].
The partial derivative ∆kf(C) gives the (marginal) contribution of the k-th player to coalitionC. Note that the same player might have a positive contribution to some coalitions and a negative contribution to other coalitions. Such a setting can model situations where some players have conflicts, which prevents them from cooperating. The lattice derivative∨kf(C)gives the outcome if thek-th player acts benevolently and joins (or leaves) the coalitionC only if this increases the worth.
Similarly,∧kf(C)represents the outcome if thek-th player acts malevolently.
Games corresponding to locally monotone functions have the property that if two coalitions are close to each other, then any given player relates in the same way to these coalitions. More precisely,f isp-locally monotone if and only if whenever two coalitions differ in less than pplayers, then the contribution of any player is either nonnegative to both coalitions or it is nonpositive to both.
Finally, let us interpret permutability of lattice derivatives. LetP be ap-subset of [n], and let C ⊆[n]∖P. Suppose that some players of P are benevolent and some of them are malevolent, and they are asked one by one to join coalitionC if they want to. We obtain the least possible outcome if the malevolent players are asked first, and we get the greatest outcome if the benevolent players make their choices first. The function f has p-permutable lattice derivatives if and only if these extremal outcomes coincide, i.e., if the order in which the players make their choices is irrelevant for everyp-subset of[n].
4. Symmetric functions
In the previous sections we saw that both notions of local monotonicity and of permutable lattice derivatives lead to two hierarchies of pseudo-Boolean functions which are related by the fact that each p-local monotone class is contained in the corresponding class of functions havingp-permutable lattice derivatives. Now, in general this containment is strict. However, under certain assumptions (see, e.g., Theorem 12), p-local monotonicity is equivalent to p-permutability of lattice derivatives. Hence it is natural to ask for conditions under which these two notions are equivalent.
In this section we provide a partial answer to this problem by focusing on sym- metric pseudo-Boolean functions, i.e., functionsf∶Bn→Rthat are invariant under all permutations of their variables. Quite surprisingly, in this case the notions of p-local monotonicity andp-permutability of lattice derivatives become equivalent.
Symmetric functions of aritynare in a one-to-one correspondence with sequences of real numbers of length n+1, where the function corresponding to the sequence α=α0, . . . , αn is given byf(x)=α∣x∣(x∈Bn). Clearly,f is isotone if and only if the corresponding sequence is nondecreasing, i.e., α0 ≤α1 ≤⋯≤αn. Similarly, f is antitone if and only if α0 ≥α1 ≥⋯≥αn, andf is monotone if and only iff is either isotone or antitone.4
It is easy to see that if f is symmetric, then every section of f is also sym- metric; moreover, iff corresponds to the sequenceα=α0, . . . , αn, then the p-ary sections of f are precisely the symmetric functions corresponding to the subse- quences5 αi, αi+1, . . . , αi+p of α of length p+1. This observation and Theorem 8 lead to the following description ofp-locally monotone symmetric pseudo-Boolean functions.
Proposition 25. Let f∶Bn→Rbe a symmetric function corresponding to the se- quenceα=α0, . . . , αn. Thenf isp-locally monotone if and only if each subsequence of lengthp+1 ofαis either nondecreasing or nonincreasing.
Unlike in the previous sections, here it will be more convenient to discard the inessential k-th variable of the lattice derivatives ∧kf and ∨kf, and regard the
4Since iff is isotone (resp. antitone) in one variable, then it is isotone (resp. antitone) in all variables.
5Here by a subsequence we mean a sequence of consecutive entries of the original sequence.
latter as (n−1)-ary functions. Clearly, if f is symmetric, then so are its lattice derivatives. Moreover, if f corresponds to the sequenceα=α0, . . . , αn, then ∧kf and∨kf correspond to the sequences
α0∧α1, α1∧α2, . . . , αn−1∧αn and α0∨α1, α1∨α2, . . . , αn−1∨αn,
respectively, for allk∈[n]. Since these sequences do not depend onk, we will write
∧f and∨f instead of∧kf and∨kf, and we will abbreviate
±∧⋯∧
ℓ
f and ∨⋯∨
±ℓ
f
by∧ℓf and∨ℓf, respectively.
The next example shows that Theorem 19 cannot be sharpened.
Example 26. Let f∶Bn → B be the symmetric function corresponding to the sequence
α = 0,0,
³¹¹¹¹¹¹¹¹¹¹·¹¹¹¹¹¹¹¹¹µp
1, . . . ,1,
³¹¹¹¹¹¹¹¹¹¹·¹¹¹¹¹¹¹¹¹µp
0, . . . ,0,1,1
wheren=2p+4 andp≥2. It follows from Proposition 25 thatf is exactlyp-locally monotone. To compute∧f, it is handy to construct a table whose first row contains the sequenceα, and in the second row we writeαi∧αi+1 betweenαi andαi+1:
f∶ 0 0 1 1 ⋯ 1 1 0 0 ⋯ 0 0 1 1
∧f∶ 0 0 1 1⋯1 1 0 0 0 ⋯0 0 0 1 Thus∧f corresponds to the sequence
0,0,
p−1
³¹¹¹¹¹¹¹¹¹¹·¹¹¹¹¹¹¹¹¹µ 1, . . . ,1,
p+1
³¹¹¹¹¹¹¹¹¹¹·¹¹¹¹¹¹¹¹¹µ 0, . . . ,0,1,
and a similar calculation yields that∨f corresponds to the sequence 0,
p+1
³¹¹¹¹¹¹¹¹¹¹·¹¹¹¹¹¹¹¹¹µ 1, . . . ,1,
p−1
³¹¹¹¹¹¹¹¹¹¹·¹¹¹¹¹¹¹¹¹µ 0, . . . ,0,1,1.
Now Proposition 25 shows that∧f and∨f are exactly(p−1)-locally monotone.
Remark 2. Example 26 shows that the degree of local monotonicity can decrease, when taking lattice derivatives, and Theorem 19 states that it can decrease by at most one. Other examples can be found to illustrate the cases when this degree stays the same, or even increases. For instance, consider the functionf(x)=x1⊕ ⋯⊕xn, which is not even 2-locally monotone, but its lattice derivatives are constant.
We conclude this section by proving that for symmetric functions the notions of p-local monotonicity andp-permutability of lattice derivatives coincide.
Theorem 27. Iff∶Bn→Ris symmetric, thenf isp-locally monotone if and only if f has p-permutable lattice derivatives.
Proof. By Theorem 21, it is enough to show that if a symmetric functionf∶Bn→R is not p-locally monotone, then it does not have p-permutable lattice derivatives.
So suppose that f is a symmetric function which is not p-locally monotone, and which corresponds to the sequence α= α0, . . . , αn. Let αi, . . . , αi+ℓ be a shortest subsequence ofαthat is neither nondecreasing nor nonincreasing. Proposition 25
implies that there is indeed such a subsequence for ℓ+1 ≤p+1. From the mini- mality ofℓit follows that the subsequence αi, . . . , αi+ℓ−1 is either nondecreasing or nonincreasing. We may assume without loss of generality that the first case holds;
the second case is the dual of the first one. Then we must have αi+ℓ−1 > αi+ℓ, since otherwise the whole subsequenceαi, . . . , αi+ℓ would be nondecreasing. Thus we have the following inequalities:
(13) αi ≤ αi+1 ≤ ⋯ ≤ αi+ℓ−1 > αi+ℓ.
From the minimality ofℓ, we can also conclude that the subsequenceαi+1, . . . , αi+ℓis either nondecreasing or nonincreasing. Asαi+ℓ−1>αi+ℓ, the first case is impossible, thereforeαi+1, . . . , αi+ℓis nonincreasing, and we must haveαi<αi+1since otherwise the whole subsequenceαi, . . . , αi+ℓwould be nonincreasing:
(14) αi < αi+1 ≥ ⋯ ≥ αi+ℓ−1 ≥ αi+ℓ. Comparing (13) and (14), we obtain
αi < αi+1 = ⋯ = αi+ℓ−1 > αi+ℓ.
To simplify notation, we setβ∶=αi, γ∶=αi+1, δ∶=αi+ℓ. With this notation we have that αcontains the subsequence β, γ, . . . , γ, δ of length ℓ+1 withβ, δ <γ. In the following we will use this observation to prove thatf does not haveℓ-permutable lattice derivatives.
Let us compute the sequence corresponding to∨∧ℓ−1f. We can construct a table as in Example 26, but this time the table has ℓ+1 rows (in the last rowµstands forβ∨δ):
f ∶ α0 ⋯ β γ γ γ ⋯ γ γ γ δ ⋯ αn
∧f ∶ ⋯ β γ γ ⋯ γ γ δ ⋯
∧2f ∶ ⋯ β γ ⋯ γ δ ⋯
⋯ ⋯ ⋯ ⋯ ⋯ ⋯ ⋯
∧ℓ−2f ∶ ⋯ β γ δ ⋯
∧ℓ−1f ∶ ⋯ β δ ⋯
∨ ∧ℓ−1f ∶ ⋯ µ ⋯
A similar table can be constructed for∧ℓ−1∨f:
f ∶ α0 ⋯ β γ γ γ ⋯ γ γ γ δ ⋯ αn
∨f ∶ ⋯ γ γ γ ⋯ γ γ γ ⋯
∧ ∨f ∶ ⋯ γ γ ⋯ γ γ ⋯
⋯ ⋯ ⋯ ⋯ ⋯ ⋯ ⋯
∧ℓ−3∨f ∶ ⋯ γ γ γ ⋯
∧ℓ−2∨f ∶ ⋯ γ γ ⋯
∧ℓ−1∨f ∶ ⋯ γ ⋯
Since β, δ < γ, we have µ < γ, and this means that the sequences corresponding to ∨ ∧ℓ−1f and ∧ℓ−1∨f differ in at least one position, therefore f does not have ℓ-permutable lattice derivatives. As ℓ ≤ p, this implies that f does not have p- permutable lattice derivatives either, according to Theorem 23.
Remark 3. As a consequence of Theorem 27, we can observe that any exactly p-locally monotone symmetric function (for instance, the functions considered in Example 26) hasp-permutable but not(p+1)-permutable lattice derivatives.
